G12 · Design YouTube Recommendations G12 · 设计 YouTube 推荐
Verified source经核实出处
YouTube two-tower recommender is public (Covington 2016). Credibility A.
Key decisions关键决策
- **Candidate generation**: two-tower dense embedding, ANN (ScaNN) returns top-K in ms.**召回**：two-tower 嵌入 + ANN（ScaNN）毫秒级取 top-K。
- **Ranking**: multi-task deep model — click + watch-time + satisfaction; weighted objective.**排序**：多任务深度模型——点击 + 观看时长 + 满意度；加权目标。
- **Freshness branch**: boost videos < 24 h; cold-start uses metadata / channel priors.**新鲜度分支**：< 24 h 视频加权；冷启用元数据/频道先验。
- **Diversity reranker**: MMR to avoid 5 videos from same channel.**多样性 re-rank**：MMR 避免同频道连发 5 条。
Follow-ups追问
- A/B infra? interleaving for sensitive metrics; separate serving shard.A/B 基础设施？敏感指标用 interleaving；独立服务 shard。